Research map / Detection and segmentation / Open-set and few-shot detection
X-Paste: Revisiting Scalable Copy-Paste for Instance Segmentation using CLIP and StableDiffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add X-Paste (CLIP-filtered real + StableDiffusion generated instances) as Copy-Paste augmentation | CenterNet2 Swin-L · LVIS 1.0 val | AP mask (all) and AP rare mask: +2.1 mask AP; +6.5 rare mask AP | Helped |
| Add X-Paste (150k generated + 150k real instances) to CenterNet2 ResNet50 | CenterNet2 ResNet50 · LVIS val (ResNet50 setting) | AP box / AP mask: +2.1 box AP; +1.9 mask AP over baseline | Helped |
| Use X-Paste instances vs Copy-Paste augmentation | CenterNet2 Swin-L · LVIS 1.0 val | AP mask (all) and AP rare mask: +0.7 mask AP over Copy-Paste; +4.8 rare mask AP | Helped |
| Combine X-Paste with Copy-Paste | CenterNet2 Swin-L · LVIS 1.0 val | AP box / AP mask: 50.9 box, 45.4 mask (vs 50.1/44.4 X-Paste alone) | Helped |
| Add X-Paste to CenterNet2 Swin-L | CenterNet2 Swin-L · COCO val | AP box / AP mask: +1.3 box AP; +0.9 mask AP | Helped |
| Add X-Paste to CenterNet2 ResNet50 | CenterNet2 ResNet50 · COCO val | AP box / AP mask: +0.6 box AP; +0.1 mask AP | Helped |
| Add X-Paste (open-vocab setting, novel rare classes withheld from training) | Detic box-supervised init CenterNet2 · LVIS open-vocabulary novel | AP novel mask: 21.4 vs 16.4 box-supervised baseline | Helped |
| Increase number of generated+real instances from 100k to 300k each (ablation) | X-Paste with ResNet50 · LVIS val | AP mask: 32.3 -> 32.9 mask AP | Helped |
| Use only CLIP-filtered real instances vs only generated vs both (ablation) | X-Paste instance source · LVIS val | AP mask: Gen only 31.6, real only 32.0, both 32.7 | Mixed |
| Adding StableDiffusion-generated images to one million retrieved images (ablation) | Swin-L X-Paste · LVIS val | AP mask: 43.5 -> 44.4 mask AP (Swin-L); 32.3 -> 33.0 (ResNet50) | Helped |
| CLIP-guided selection of mask among four segmenters (vs single segmenter) (ablation) | instance mask generation · LVIS val | AP mask: max CLIP 31.8 vs U2Net 31.7, UFO 31.7, CLIPseg 31.6; rare AP lower than UFO | No clear effect |
| Apply CLIP score threshold 0.21 instead of no filtering (ablation) | instance filtering · LVIS val | AP mask / rare AP mask: +0.6 mask AP; +1.4 rare mask AP | Helped |
Closest papers by meaning
- The Devil is in the Object Boundary: Towards Annotation-free Instance Segmentation using Foundation Models (ICLR 2024 · similarity 0.87)
- InstanceDiffusion: Instance-level Control for Image Generation (CVPR 2024 · similarity 0.86)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.85)
- MaskDiff: Modeling Mask Distribution with Diffusion Probabilistic Model for Few-Shot Instance Segmentation (AAAI 2024 · similarity 0.85)
- Prompting Diffusion Models for Zero-Shot Instance Segmentation (NeurIPS 2026 · similarity 0.85)
- Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.